Wildly Uncensored Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF

    
        By vramkickedin    
     | 
    
            June 24, 2026 at 6:10 pm        
    
     | 
    
        2 min read

A new uncensored model is now available on Hugging Face and it carries the long name Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF. This release is a 40-billion-parameter dense language model that was expanded from Qwen's 27B base and then trained with high-reasoning datasets using Unsloth AI on local hardware. The project combines several modifications, including uncensored filtering through the "Heretic" process, character and intelligence tuning via "Deckard" datasets, and final training on a Claude 4.6 Opus distill to improve reasoning stability and efficiency.

DavidAU who is also known for Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF, built this model as a fully uncensored, high-performance reasoning system. The creator warns that the model has "character" and will provide graphic or NSFW content when prompted, having been stripped of all safety alignment. It was engineered for people who want a model with strong personality, sharp intelligence, and zero built-in filtering.

Balanced quant engineering for maximum precision

Key features of this release

Dual imatrix dataset used for quant optimization.
Quant scores reach 98% of full BF16 precision.
96 layers and 1,275 tensors for deeper processing.
Variable length reasoning adapts to task complexity.
256K native context window for long conversations.
Exceeds the base model on six of seven benchmarks.
Fully uncensored with no safety alignment remaining.
Suggested settings included for thinking and coding tasks.

This model targets users who want a powerful unfiltered assistant for creative writing, deep reasoning, or complex coding tasks. The quant engineering focused on balance and precision rather than raw power, helping the model stay stable during long conversations and tool-calling scenarios.

What the developer wants you to know

The creator notes that dual imatrix datasets, combining NEO and NEO-CODE sources, were used to measure and merge the best quantization statistics. Every quant was individually benchmarked against the full BF16 model, and specific tensor adjustments were calibrated through testing to ensure long-context and multi-turn conversation performance. Lower quants may experience looping with weak prompts, but adding a system prompt or slightly adjusting repetition penalty usually resolves the issue.